AI can predict when complex systems like forests, animal populations or the power grid will suddenly start behaving very differently. Identifying such tipping points may help avert disastrous collapses in biology or human infrastructure.
鈥淗istory is full of harmful critical transitions, such as financial market crashes, disease outbreaks and blackouts,鈥� says at Tongji University in China.
To make predicting such transitions more precise, he and his colleagues combined two different types of artificial intelligence called neural networks. They optimised the first one to understand the functioning of and connections across systems structured like large networks with many nodes. For example, in an ecosystem, each node would be a geographical location where researchers would collect data about how many animals or trees live there. Nodes could also be different parts of the power grid or areas where disease outbreaks are occurring.
Advertisement
The team designed the second neural network to be especially good at analysing how networks change over time. So, the first network would process data about each node and the interactions between them, then feed into the second network, which detected patterns in data that recur over time and predicted future tipping points.
Yan says that past studies focused on identifying particular features of data that increased or decreased as a tipping point approached, but his team鈥檚 AI goes further. 鈥淚t aims to pinpoint the precise conditions that lead to system collapse, asserting: 鈥榃atch out, if the system reaches this [specific] condition, it will collapse immediately鈥�,鈥� he says.
He and his colleagues tested the AI on a range of mathematical models and simulated data used to represent power grids, crop harvests and animal populations. In one test, they used real-world data on vegetation and rainfall in a forest ecosystem in Central Africa that suddenly became a savannah. The researchers trained the AI on simulations and the scarce data available for one part of the region, and then had it predict the value for annual precipitation at the tipping point for another. The AI correctly predicted what had actually happened to the ecosystem, even when it was only given data for about 10 per cent of the nodes to learn from.
at Case Western Reserve University in Ohio says the new approach is 鈥減owerful鈥�, even though machine learning doesn鈥檛 offer the same level of insight into why a tipping point occurs as a full-fledged mathematical model might. But the advantage of AI methods is that they can be applied to many systems instead of having to be formulated for a specific one each time, and they can be optimised to deal with incomplete or sparse datasets, she says. 鈥淲e really, really need both. Machine learning is telling us how to get more clues out of data, and theory is telling us what to do with those clues,鈥� says Abbott.
Now, Yan and his colleagues want to apply their method to even more systems like floods, wildfires and disease outbreaks. And they are interested in not only when but also why tipping points happen. 鈥淲e aim to delve deeper into the algorithm鈥檚 black-box nature to explicitly uncover the specific features used for predictions,鈥� he says.
Physical Review X
